← /editions
#73

Stop trusting the model to say no

// A new paper argues prompt injection is an authorization architecture problem rather than a model problem, and backs it with compromised-model evaluations showing AgentDojo exfiltration dropping to zero. Anthropic and OpenAI shipped adjacent containment and supervision infrastructure the same week.

$ today_in_10_seconds

A new paper argues prompt injection isn't a model problem at all — it's an authorization architecture problem, and it has the eval numbers to make the case uncomfortable. Meanwhile Anthropic shipped egress allowlists for its managed agents and OpenAI previewed cross-session abuse detection that never shows staff your content. The constraint layer is moving outside the model. That's the story.

$ lead_story

Bounded Agents: Delegation Security for Multi-Agent AI Systems

Xabier Muruaga's paper starts from an observation most agent stacks quietly violate: permissions are set at session start, then every request is judged in isolation, with no memory of what the agent already did. His Agentic Principal Chain tracks delegated authority as it passes from principal to sub-agent, narrowing scope and budget at each hop, and runs six authorization checks against accumulated session state rather than against the request alone. The headline result comes from a compromised-model evaluation — the attack tool call is injected as ground truth after the first legitimate one, so the model's judgement is taken out of the equation entirely. Across 3,154 instances spanning InjecAgent, AgentDojo and ASB, AgentDojo exfiltration went from 75–100% to zero in all four domains, all 544 InjecAgent data-stealing cases were blocked, and destruction dropped from 38.6% to 4.0%. Authorization cost 0.24ms at p99. Utility cost 8.6 and 13.9 percentage points. Caveats worth holding: single author, not peer reviewed, self-run evals, and the soundness proof only covers prohibited combinations under a complete restriction set with serialized admission — which is a real constraint if you're running agents in parallel.

// why_it_matters: If the enforcement point sits outside the model, a compromised model stops being a catastrophic failure and starts being a blocked request.

read more →

$ also_worth_attention

  • Muxel — a multi-agent terminal multiplexer

    A native GPL-3.0 terminal app built on GPUI that treats a fleet of coding agents as the primary object: splittable panes with live working/idle/blocked/done status, one git worktree per agent, Ctrl+Shift+A to jump to whichever agent is waiting on you, and scheduled loops that fire a saved prompt on a cron. Presets ship for Claude, opencode, Amp, Grok, Ollama and others. The feature list is the project's own; what's notable is the design assumption underneath it — that the scarce resource is no longer agent throughput but your attention, and the interface should optimise for the 'needs you' counter.

    link →
  • Anthropic ships domain allowlists for Managed Agents

    This week's Claude Developer Platform release lets you constrain which sites a managed agent's web_search and web_fetch can reach via allowed_domains / blocked_domains, adds memory stores that sync into self-hosted sandboxes, and rebuilds the Console session viewer around per-tool statistics, mounted resources and per-session cost. Managed Agents is still beta behind a header, so treat the surface as unstable. But egress control plus a per-tool audit view is precisely the 'enforce it outside the model' pattern the lead story formalises — arriving as product the same week it arrives as theory.

    link →
  • OpenAI previews Private Safety Processing

    Announced 19 August: a safety system designed to spot misuse patterns across related interactions rather than judging each one alone, while keeping Zero Data Retention intact and giving OpenAI staff no access to the underlying content. The motivation is explicitly agentic — long autonomous tasks hide risk that single-turn evaluation can't see. Rollout and a technical white paper are slated for September, and until that paper lands the central claim is unverifiable from the outside. One footnote survives ZDR regardless: images flagged as potential CSAM are still retained for manual review and legally required reporting.

    link →

$ bigger_picture

The trend behind the headlines

Four items, one direction of travel: the interesting work has moved off the model and onto everything wrapped around it. Muruaga's paper makes the strongest version of the argument — that a prompt injection is only dangerous if the agent held the authority to act on it, which makes this an authz design question rather than an alignment one. Anthropic is shipping the crude version of that today as domain allowlists and per-tool audit. OpenAI is solving the adjacent problem of how you supervise multi-step work without holding the data. Muxel is solving the human end: which of your twelve agents is blocked right now. Note the price tag in the paper, though — roughly 9 to 14 points of utility for that containment. Anyone selling you safety at zero cost to capability hasn't measured it.

$ tool_of_the_day

AgentDojo

ETH Zurich SpyLab's benchmark for prompt injection attacks and defenses against tool-calling agents, and the environment today's lead story reports its cleanest numbers on. It pairs realistic user tasks with injection tasks across banking, Slack, travel and workspace suites, and its plugin architecture lets you register your own defense at the tool-filter or output-processing layer and get a comparable score rather than a vibe. If you're evaluating a runtime guardrail vendor, ask which AgentDojo configuration they ran and what it cost them in utility.

explore AgentDojo →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime